Course contentsShow
← AI Engineering
Lesson 1577 of 1,88638. Bias, Fairness, and AlignmentFree lesson

Post-processing: Output Calibration

Adjust model predictions after training to satisfy fairness criteria while maintaining overall performance metrics.

Post-processing: Output Calibration

What you'll learn: How to adjust model predictions after training to meet fairness criteria without retraining the model.

The Core Idea

Sometimes you've already trained your model, and retraining isn't practical or possible. Post-processing techniques let you modify the model's outputs—its predictions or scores—to satisfy fairness constraints you discovered through bias detection and fairness metrics evaluation.

Think of it like color-correcting a photograph after it's taken. The original image (model) remains unchanged, but you adjust the final output to fix imbalances while keeping the overall quality acceptable.

How Output Calibration Works

Output calibration adjusts prediction scores or decision thresholds for different demographic groups to achieve fairness goals like equalized odds or demographic parity. For example:

  • If your model predicts loan approval scores, you might adjust the threshold for different groups so that qualified applicants have equal acceptance rates across demographics
  • In text generation, you might post-process outputs to reduce stereotypical associations detected by fairness metrics
  • For classification tasks, you can recalibrate confidence scores to ensure predictive parity across groups

The key advantage: you don't need access to training data or model internals. You only need the outputs and sensitive attribute labels from your test or validation set.

The Trade-off

Like all fairness interventions, output calibration involves trade-offs. Adjusting thresholds or scores for fairness may slightly reduce overall accuracy or other performance metrics. You're balancing competing objectives: individual prediction quality versus group-level fairness guarantees.

Key Takeaway: Post-processing output calibration lets you enforce fairness constraints by adjusting model predictions after training, offering a flexible intervention that doesn't require model retraining but may impact overall performance metrics.